Docs Lookup

Docs Lookup (with fallback)

When you need authoritative information about an unfamiliar tool / API / error message / library behavior, follow this fallback chain. Each step has explicit success and failure criteria so the agent never silently degrades to guessing.

Step 1: Structured docs MCP   (Context7; or Microsoft Learn for .NET/Azure/Windows/MSVC)
   ↓ (unavailable / no match / wrong content / tool error)
Step 2: WebSearch     (broader web, recent ecosystem changes)
   ↓ (unavailable / no useful results / paywalled)
Step 3: Calibrated uncertainty (tell user what you tried + what's needed)

You do not skip steps. If Step 1 fails, you go to Step 2. If Step 2 fails, you go to Step 3. You do NOT retry Step 1 with the same query hoping for a different answer.

Step 1: Structured docs MCP — Context7, or a first-party MCP for its ecosystem

When it's the right tool: API references, configuration options, well-known libraries (React, CMake, Clang, etc.), version-specific behavior.

Two-call usage (typical Context7 / Microsoft Docs MCP shapes):

1. resolve-library-id(library_name)            → canonical id, e.g. /llvm/clang
2. get-library-docs(id, topic="<question>")    → structured snippet + citation URL

(Tool names vary by harness: mcp__Context7__*, mcp__plugin_context7_context7__*, mcp__73bcda44-…__query-docs, etc. Use whatever resolves to "Context7" in your tool inventory.)

First-party ecosystem docs MCP — prefer it over Context7 when the topic matches. Context7 is the general default for libraries, but if the question is about a specific vendor ecosystem and that vendor's docs MCP is in your inventory, it is higher-signal (first-party). The prime example: Microsoft Learn (microsoft_docs_search → microsoft_docs_fetch; microsoft_code_sample_search for snippets) for .NET / Azure / Windows / MSVC / Win32 topics — including MSVC compiler errors and Windows-SDK headers in C/C++ work. Same two-call shape (search → fetch) and the same failure-mode handling below; if it is absent or returns nothing useful, fall through to Context7 → WebSearch as normal. (General principle — prefer the first-party MCP for its ecosystem; not a hard dependency.)

Context7 "doesn't work" — what to watch for:

Do not retry the same query multiple times — Context7 is deterministic for a given input. One retry with a refined query is allowed; further iteration is the agent thrashing.

Step 2: WebSearch (fallback)

When it's the right tool:

• Context7 unavailable or returned nothing useful
• Question is about a recent change (release notes, CVEs, ecosystem shifts)
• The exact error string is likely indexed by search engines (compiler errors, framework stack traces)
• The library is small / non-indexed in Context7

Query construction rules:

Source quality (descending trust):

1. Official docs (linked from the tool's homepage)
2. Stack Overflow answer with accepted check + multiple upvotes
3. GitHub issue with maintainer reply
4. Reputable engineering blog (Anthropic, OpenAI, Cloudflare, etc.)
5. Random blog posts — only when corroborated by 1-4

WebSearch "doesn't work":

Step 3: Calibrated uncertainty (terminal fallback)

When external lookup fails, do not guess. State explicitly what you tried and what you'd need:

"I looked up <topic> via Context7 (not loaded / returned no match for <query>) and WebSearch (no useful results / returned only AI-generated content). I am uncertain about <specific question>. To proceed I would need: <the specific input from you, e.g. output of \cmake --version`, or a link to the right docs, or permission to try option X cautiously>`."

This earns user trust. Guessing destroys it. The terminal fallback exists because pretending to know is the highest-cost agent failure mode.

Reporting what you looked up

When you DO find an answer, cite the source:

"Per Context7 (clang-tidy v18 docs): bugprone-easily-swappable-parameters flags adjacent params of the same type that could be confused at call sites. Tuning option: MinimumLength (default 2)."

"Per WebSearch (Stack Overflow answer with 47 upvotes, accepted Aug 2025): the fix is set(CMAKE_POLICY_DEFAULT_CMP0148 NEW) before find_package(Python)."

The citation makes your claim falsifiable — the user can verify. Unsourced confidence is worse than sourced uncertainty.

When NOT to invoke docs-lookup

• The answer is in the project's own docs (use project-indexing to find local files first)
• The question is about the user's intent (ask them directly)
• You already have high-confidence evidence from a prior lookup this session (don't re-look-up the same thing)
• The question is project-internal architecture (read the code; don't search the web)

Related

• self-correction-loop — when a tool error needs context, this skill is the lookup step
• anti-hallucination-gates — the citation from this skill becomes part of evidence
• project-indexing — for internal lookups (when info is in your repo, not on the web)
• All cpp-* skills — domain entry points; they delegate the lookup procedure here